基于机器学习的新生儿坏死性小肠结肠炎的鉴别诊断

Differential Diagnosis of Neonatal Necrotizing Enterocolitis Based on Machine Learning

坏死性小肠结肠炎(Neonatal Necrotizing Enterocolitis, NEC)是一种很常见的、最严重的胃肠道急症,尤其是新生儿和极低体重儿,发病率和死亡率极高,早期NEC的临床症状以及放射学等多表现为非特异性,因此NEC的鉴别诊断尤为重要。搜集了广州市妇女儿童医疗中心2011—2018年的248例患者。通过递归式消除特征的方法(RFECV_RF)选择最优的特征集,运用XGBoost、决策树、人工神经网络3种不同的方法进行训练,并用五折交叉验证以及超参数搜索选择最优参数,构建最优模型,用这3种方法得到的分类器对50例临床未能确诊NEC的病患进行预测。实验结果表明,在50例测试样本中,XGBoost敏感性80.48%,特异性100.00%,AUC90.24%;决策树敏感性60.98%,特异性82.93%,AUC71.95%;人工神经网络敏感性78.05%,特异性100.00%,AUC89.02%。采用机器学习的方法为NEC的鉴别诊断提出了一种新的研究思路和方法,具有很重要的临床价值。

Neonatal necrotizing enterocolitis(NEC)is a common and most serious gastrointestinal emergency,especially in neonates and very low birthweight infants.NEC is extremely prevalent in morbidity and mortality.The clinical symptoms and radiology of early NEC are mostly non-specific,so the differential diagnosis of NEC is particularly important.This paper collected 248 patients from Guangzhou Women and Children Medical Center from 2011 to 2018.Eliminated by recursion feature method(RFECV_RF)to select the optimal feature set,by XGBoost,decision tree,artificial neural network of three different methods for training,and five cross validation and super parameter search optimal parameters choice,build the optimal model,using the three methods of classifier for50 cases of clinical predict failed to diagnose NEC patients.The results showed that XGBoost sensitivity 80.48%,specificity 100.00%and AUC 90.24%in the 50 test samples.The sensitivity,specificity and AUC of the decision tree were 60.98%,82.93%and 71.95%respectively.The sensitivity,specificity and AUC of ANN were 78.05%,100.00%and 89.02%respectively.This paper presents a new method for NEC differential diagnosis by machine learning,which is of great clinical value.